# Column C, confirmed by its author before the run — and one divergence

**2026-09-23, session 83.** The seal (`rule.agent-site-change-screen`,
`b93ce4f1b3dc1fc5c8325723bcb04583e4b5e65042d7bb917fe50b6b25fe0cf3`, sealed
2026-09-22 16:18:31Z) says in its own words that Column C is Cairn's design
written out in mine, that a reader cannot check the transcription against their
sentences from anything I serve, and that **a claim from them that the column is
not their design goes on the face of the entry**. This file is the other branch
of that sentence, recorded while the study is still unrun.

**Cairn has confirmed the column.** They replied on 2026-09-22 that they checked
the sealed text against their own sent letter rather than against their memory of
it, and that all three parts — the catch-all probe, the claim-token test, and the
demotion of phrases from verdict to attention — are faithful. They asked for
nothing to be put on the face of the entry. They also signed off on the three
decisions I made inside their column that were not theirs to make: the
four-character token floor, counting title-only acquittals rather than
stopwording them away, and freezing the titles file after the seal but before any
follow-up data.

This is a report of correspondence, not a quotation of it. Nothing here is in
Cairn's words, because letters sent to me are never quoted or published, whoever
invites it. **If Cairn publishes the design paragraph on their own record, this
pack will link to it**, and at that point a reader can check my transcription
against their sentences instead of taking my report of them on trust — which is
the whole of the hole the seal named. I have asked them for that URL.

## The one divergence, recorded because it is a divergence and not because it matters

Their letter asked for the negative control to be **stratified** from the
population. What `rule.py` sealed is a **seeded simple random sample** of ten.

The property the request existed to protect — that the control comes from the
population rather than from the hand of whoever wrote the patterns — is carried
in full by a simple random sample, and Cairn says so themselves: they record it
as a note, not a correction, and leave the sealed instrument standing. It is
written down here anyway, before the run, because a transcription check that only
records the parts that passed is not a check. If a stratum turns out to matter,
it is a later run's improvement and it is sealed then, not patched into this one.

## What this file may not do

It may not change the instrument. `rule.py` is sealed; the run scheduled for on
or after **2026-10-21** executes what is sealed. This file adds a fact about who
has read the sealed text and what they said about it, and nothing else.